Service Units & Budgets

From Fall 2026, GPU time on PARASOL (formerly DSMLP) is metered in Service Units (SU), and each workspace member has a budget of them.

Default budget

10 SU per week per member, in every workspace

Renews

Each Monday at 00:00 Pacific time, weekly by default

Off-peak

50% off, 1 AM to 4 PM Pacific time

Free to cancel

At least 24 hours before the start

What a Service Unit Is¶

A Service Unit (SU) is a usage credit for GPU time. Each GPU class carries an hourly SU rate, and every reservation is priced from that rate. The scheme follows the HPC allocation model: a budget of credits, drawn down as GPU time is used and renewed on a cycle. Coming from HPC lists HPC terms and their PARASOL equivalents.

Service Units divide capacity and are not billed. Storage is not measured in Service Units. Workspace and Personal Quotas covers storage charges.

What Spends Service Units¶

On-Demand Lease Charges¶

Launching a GPU session without a booking creates an on-demand lease: a reservation made on the spot, which draws Service Units exactly as a booked window would. This applies whether the reservation app is ever opened.

A lease covers the session's declared runtime plus 10 minutes, and is charged when it is granted. A session that declares no runtime is given 1 hour, so it holds a lease of 1 hour 10 minutes. The declared runtime is separate from the 6-hour runtime limit, which stops the pod. When the session ends, the lease is cancelled and the charge is recalculated:

• The time used is charged in full.
• Unused time in the lease's first 2 hours is free.
• Unused time after the first 2 hours is charged on the scale in The Cancellation Penalty.

A lease of 1 hour 10 minutes ends inside its first 2 hours, so it is charged for the time used and nothing more. For example, a default launch on one medium GPU at the full rate is charged about 1.17 SU when the lease is granted. If the session stops after 40 minutes, about 0.67 SU is kept.

A session that runs past the end of its lease draws no further Service Units, and is no longer protected; see Overstay. A longer guarantee comes from Extend, which is a booking and is charged like one.

• An unused session is charged at the same rate as a used one. The charge is for holding the GPU, not for the work done on it.
• A session that stops using its GPU continues to draw Service Units until idle culling reclaims the GPU. Idle culling waits longer on a quiet cluster than on a busy one. The Timings lists the thresholds.
• A script that starts a new GPU session on each iteration, or a job restarted repeatedly during debugging, is charged on every launch. Nothing stops the repeated charges.
• A pod left waiting for a lease starts, and is charged, whenever the lease is granted. See Waiting for an On-Demand Lease.
• A pod limited with -n or -v to a node or GPU model that has no free GPU is given no lease, and is not charged, until a node it allows has one. Under a booking, the same pod holds the booking, which is already charged, while it waits. See Node Selection.

Limiting the Spend¶

Stop sessions that are not in use. From the browser, use File → Hub Control Panel → Stop My Server. From a terminal, run kubectl delete pod <pod-id>. Stopping a Session describes the browser procedure.

Request the smallest GPU class the work fits in. A larger class is not faster for a model that fits in a smaller one, and it costs more per hour; see What an Hour Costs. Choosing a Class covers class selection.

Check the remaining balance before a long run, not after it; see Remaining Balance.

A booked window costs no more than an on-demand lease for the same hours. Reservations describes booking. Off-peak hours are discounted. See Peak & Off-Peak Hours.

The Size of a Budget¶

Every workspace gives each member 10 SU per week by default, in course and research workspaces alike, and in ORG_ON_DEMAND, the default workspace. The instructor or researcher can ask for a different figure. Changing a Budget lists the routes for changing a budget.

At the rates in What an Hour Costs, 10 SU buys 10 GPU-hours of medium at the full rate, or 20 off-peak.

A budget belongs to a workspace, not to a person. A person who is a student in one course, a TA in another, and a member of a lab holds a separate budget in each, on that workspace's cycle. A reservation draws on the budget of the workspace it is made under: the workspace chosen in the booking wizard, or, for an on-demand lease, the workspace the pod's dsmlp/course label names. launch.sh -W sets that label, and a pod that names no workspace is charged to ORG_ON_DEMAND. See The Default Workspace and Belonging to Several Workspaces.

Budgets and Workspace Quotas¶

A budget limits how much GPU time a member may spend. A workspace's quota limits how many GPUs of a class the workspace may hold at once. See Quotas, Cohorts & Availability. Either can stop a launch, for different reasons. Remaining budget does not mean a GPU is free.

A budget window and a workspace quota are independent. A quota can change from week to week, for example during a deadline surge, without any change to the budget.

What an Hour Costs¶

A reservation costs the class's rate, times the hours, times the number of GPUs. The rates in force for Fall 2026:

For example, 3 hours of two medium GPUs at the full rate costs 6 SU. The Rates page in the reservation app is the authority on current rates; see The Rates Page. The booking wizard shows a booking's cost before it is confirmed.

Peak & Off-Peak Hours¶

Service Unit rates are discounted by 50% in the off-peak hours, 1 AM to 4 PM Pacific time, on every day of the week, weekends included. The peak hours run from 4 PM to 1 AM. The discounted hours are marked on the Rates page, which shows them for each day of the week, and by an amber strip on the booking wizard's timeline. An off-peak window is also more likely to be available. On a busy deadline evening there is a walk-up queue; at midday there usually is not.

Batch and background jobs, which run unattended, are the work most easily moved off-peak, for example by starting them overnight. An interactive debugging session is not easily moved. See Job Modes. Checkpointing & Logging Long Runs covers long unattended runs.

Availability that reads zero in the evening often reads differently in the morning, particularly for a workspace in an overcommitted cohort. See Cohorts.

An off-peak booking is otherwise identical to a peak booking. The GPU hardware is the same, and the runtime guarantee, The Claim Window, and The Cancellation Penalty apply exactly as they do at peak.

The Rates Page¶

Rates, in the reservation app's sidebar, charts the rate of every GPU class for each of the 168 hours of a week. Bar height is the rate; shorter amber bars are discounted hours. Point at a bar for its exact rate, and use ‹ and › to move between weeks.

Show rate summary table lists each class's discount windows. The table can list a discount window that does not apply in the week shown; the chart is exact. The Rates page gives rates in SU/hr and the booking wizard in SU/GPU·h. Both are per GPU.

Budget Windows & Cadences¶

A Service Unit budget renews weekly by default, in every workspace: from Monday 00:00 to the next Monday 00:00, Pacific time. A workspace can be given a different window on request. The Dashboard shows the window in force; see Anchor Modes.

A reservation is charged wholly to the window its start falls in, however far it runs into the next. A booking made for a later week draws on that week's budget, not the current one's.

• Unused budget does not carry over to the next window.
• No warning is sent as a budget runs low.
• A reservation is priced when it is booked. A window that renews during a job does not change that price.

Whether a course budget continues to renew after the instructional term is not yet published. One Additional Quarter describes access after the term.

Anchor Modes¶

The Dashboard labels each workspace's budget with its mode:

Remaining Balance¶

The SU Budget card on the reservation app's Dashboard shows one block for each workspace that has a budget. Current Window gives the window's first day and the day the budget renews. A bar shows the SU committed against the budget, and a line below it gives the figures:

The bar and the figure beside it leave out Cancelled, so they do not add up to Remaining when a penalty has been kept. In team mode the block reads Team budget, and a workspace with a pool shows a Group pool row. An increase an administrator applies changes Budget without any other notice.

The booking wizard also shows a budget meter above its timeline. There, "used" means everything charged to that window: ended, open, and kept penalties.

The Group Pool¶

A workspace may hold a pool budget alongside its members' own budgets. A reservation draws on both. The pool also limits workspace managers, and can refuse a booking when the member's own budget still has room. Pool budgets are expected to be uncommon in course workspaces. The PI of a research group can state whether the group has one configured.

Changing a Budget¶

Budgets are set administratively. A workspace manager (an instructor, TA, or PI) may not edit any budget, including that of the workspace they manage. Managing a Group lists manager privileges.

To change a budget, the manager requests it by ticket to parasol@ucsd.edu, stating the workspace, what the work is, and the date by which it matters. Administrative Requests describes the request format.

A booking a manager makes on a member's behalf does not add to the member's budget. It is charged to the member.

The Cancellation Penalty¶

Cancelling a booking at least 24 hours before its start costs nothing. Any other cancellation keeps part of the booking's cost:

1. The time already used is charged in full.
2. The unused time that falls within the next 24 hours is counted. Unused time beyond that is free.
3. An exemption of half the used and counted time together, at most 8 hours, is free.
4. If the counted time is no more than the exemption, nothing more is charged. Otherwise the charge is the cost of the counted time, multiplied by the counted time less the exemption, divided by the counted time.

The counted time is priced at the rates of the hours it covers, off-peak discounts included. Examples on one medium GPU at the full rate, where 1 SU is one hour:

Two consequences follow. A booking cancelled before it starts keeps half the cost of its hours that fall within the next 24 hours; where more than 16 of its hours fall in that span, it keeps all but 8 hours' worth of them. Handing back the rest of a booking in progress is free only when the rest is no longer than the time already used, and no more than 8 hours.

A no-show is charged on the same scale, at the moment the booking is declared a no-show; see The Claim Window. An on-demand lease is charged on the same scale, except that unused time in its first 2 hours is free; see On-Demand Lease Charges.

A kept penalty appears on the Dashboard as Cancelled and counts against the budget window until the window renews. A waiver frees it at once. Penalties are assessed in Service Units. No money is billed.

Cancelling a Booking¶

1. Select Cancel

   On My Reservations, select Cancel on the reservation.

2. Read the cost

   The dialog states what the cancellation will cost:

   • For a booking that starts within 24 hours: Late cancellation: X SU will be charged for reserved hours within the next 24 h (part is exempt). When the cancellation costs nothing, no notice appears. When the app cannot work out the charge, the notice reads Late cancellation: an SU penalty may apply for reserved hours within the next 24 h.
   • For a booking in progress: You will be credited X SU., and a warning that pods running under the booking will be stopped.
   • For a booking more than 24 hours away: no notice, because nothing is charged.
3. Confirm

   Select Yes, Cancel to confirm.

   Cancelling stops running sessions

   Cancelling a booking in progress deletes the sessions running under it; see ReservationCancelled.

Cancelling in Advance¶

A booking cancelled at least 24 hours before its start costs nothing, and the capacity returns to the pool. Release a window as soon as it is known that it will not be used.

Having a Charge Waived¶

A workspace manager may waive a penalty: for a course, the instructor or TA; for a lab, the PI. Waiver requests go to the manager. A member cannot waive a penalty on their own cancellation.

A manager waives a penalty when cancelling a member's booking, with Waive this penalty, or afterwards, with the waive button on the cancelled reservation on Group Reservations. A penalty from a teammate's cancellation in team mode is waived afterwards in the same way. A manager's waiver clears the member's share of the penalty. Where the workspace has a pool, only an administrator clears the pool's share. See The Group Pool.

A waiver is a judgment about circumstances, such as an illness, a booking made on a member's behalf at the wrong hour, or a cluster problem during the window. It is not an automatic remedy for a forgotten booking.

Researchers working outside a course have no manager to ask. Where a penalty was assessed in circumstances that warrant relief, write to parasol@ucsd.edu and describe what happened.

When a Budget Runs Out¶

A spent budget blocks the member's bookings and on-demand leases until the window renews. A pending GPU session waits with an OnDemandLeaseDenied event that quotes the budget and names the time it renews; see Reservation Events.

A budget on the default weekly window renews each Monday. In a course, the instructor or TA can request an increase by ticket. See Changing a Budget. A booking the instructor or TA makes on the member's behalf is charged to the member, so it does not help.

In a research workspace, write to parasol@ucsd.edu, naming the work and the date by which it must be done. Where the deadline permits, waiting for the budget window to renew is also an option. See Budget Windows & Cadences.

Report a problem with this page